Literature Research

Read papers for a purpose, investigate questions, map the mechanisms for a design problem, preserve sources, and build on prior work. Keep ordinary files and a small Markdown index. The library must work with filesystem search alone: no database, embedding index, background service, mandatory citation manager, or custom application.

It serves a near-term investigation as well as a paper and does not require a paper plan, journal entry, or research-direction interview.

Read one paper

Pick the mode from the request; scan is the default.

• Scan: two lines: the one new thing in your words, and whether it matters for the user's stated work. If the paper hides its one thing, say so; that is information about the paper. Do not paraphrase the abstract.
• Extract: only the piece the user names, with what is needed to use it: the definition, the setup, the number with its unit and denominator, the proof step with its assumptions, and its location in the paper.
• Reproduce: when the user intends to build on or refute the work: what redoing it would take, the stated and unstated assumptions, what is claimed but not shown, errors with their location, and the one missing experiment a skeptical reader would want.

If the paper cannot be accessed, say so and stop; do not reconstruct it from the abstract or from memory.

When the paper's framing could change the user's approach, add a short assumption check, adapted from the "Read all the papers" and "Ignore all the papers" sections of Carlini's research essay. Name the conventions it inherits from the field, such as the metric, threat model, baselines, dataset, or an early arbitrary choice, and mark those that look unjustified. Ask what the user would have done on their problem without reading it and whether the paper's pull is earned. Name the one question it makes the user want to shout about, if any; that is where their own problem may be. A narrow extraction needs no such check.

A single read saves into an existing library or wherever the user asks; it does not start a library. Within an investigation, scan most sources, extract from the few that bear on the question, and reproduce only on request.